Background Modeling and Algorithm Fusion for Airborne Landmine Detection

摘要

Airborne mine and minefield detection has been an active topic of research in recent times. The detection process looks for local anomalies in the acquired mid-wave infrared imagery. An anomaly is any mine size image feature that is different from its immediate surrounds. The so called RX algorithm by Reed and Yu has been used extensively in airborne mine detection systems as the anomaly detector. The RX detector measures the local signal to clutter ratio assuming a zero mean uncorrelated Gaussian background. The airborne minefield detection system is expected to detect possible mines and minefields over various types of terrains, different soil conditions and at different times of the day. For most terrains conditions, the background is often correlated and inhomogeneous. This raises the question as to whether the performance can be made invariant to terrain characteristics. The primary aim behind this thesis is to model the distributions for RX statistics of false alarms under more realistic background conditions. This modeling is expected to facilitate the study of the detectability of mines and minefields in various backgrounds. By modeling various backgrounds, guidelines can be obtained for likely false alarms rates for different backgrounds. Algorithmic fusion is simple but often effective tool to derive optimum performance from multiple detection algorithms. The parameters obtained from background modeling can be use as a catalyst for algorithm fusion and thereby improve detection performance for a given set of detector algorithms.